CN107092594A - Bilingual recurrence self-encoding encoder based on figure - Google Patents
Bilingual recurrence self-encoding encoder based on figure Download PDFInfo
- Publication number
- CN107092594A CN107092594A CN201710257714.4A CN201710257714A CN107092594A CN 107092594 A CN107092594 A CN 107092594A CN 201710257714 A CN201710257714 A CN 201710257714A CN 107092594 A CN107092594 A CN 107092594A
- Authority
- CN
- China
- Prior art keywords
- phrase
- bilingual
- semantic
- language
- node
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Granted
Links
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F40/00—Handling natural language data
- G06F40/30—Semantic analysis
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F40/00—Handling natural language data
- G06F40/40—Processing or translation of natural language
- G06F40/42—Data-driven translation
- G06F40/44—Statistical methods, e.g. probability models
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Physics & Mathematics (AREA)
- Health & Medical Sciences (AREA)
- Artificial Intelligence (AREA)
- Audiology, Speech & Language Pathology (AREA)
- Computational Linguistics (AREA)
- General Health & Medical Sciences (AREA)
- General Engineering & Computer Science (AREA)
- General Physics & Mathematics (AREA)
- Probability & Statistics with Applications (AREA)
- Machine Translation (AREA)
Abstract
Bilingual recurrence self-encoding encoder based on figure, is related to the natural language processing based on deep learning.Bilingual phrase is extracted from parallel corpora as training data, the translation probability between bilingual phrase is calculated;Method based on pivot, calculates and repeats probability;Construct the semantic relation figure of bilingual phrase;Semantic relation figure based on bilingual phrase;Quantitative model object function, carries out model parameter training.Target is expressed as preferably to learn the insertion of bilingual phrase, lacks for conventional method and considers in natural language more fully semantic constraint relation, propose a kind of bilingual recurrence self-encoding encoder based on figure.Algorithm is clear and definite, clear thinking, can improve the bilingual phrase insertion acquired and represent, preferably act on natural language processing task.The semantic relation figure of bilingual phrase is constructed first, two implicit semantic constraints are defined by graph structure, is represented for learning more accurate bilingual phrase insertion, and then be preferably applied in natural language processing task, such as machine translation.
Description
Technical field
The present invention relates to the natural language processing based on deep learning, more particularly, to the bilingual recurrence own coding based on figure
Device.
Background technology
Natural language processing is an important research direction of Computer Subject artificial intelligence.It study how cause people with
Natural language can be used to carry out efficient communication between computer.It is one and melts linguistics, computer science, mathematics in one
Subject.
This invention is mainly concerned with bilingual recurrence self-encoding encoder of the structure based on figure, and uses it for bilingual phrase insertion
Represent modeling.Neutral net is that a kind of application is similar to the mathematical modeling that cerebral nerve cynapse connecting structure carries out information processing.
In recent years, the natural language processing research based on neutral net has become the main trend of the discipline development, various nerve nets
Network emerges in an endless stream.Wherein, recurrent neural network (Recursive Neural Network, referred to as RecNN) is widely used in
The embedded of text represents research.(this structure can pass through minimizes reduction mistake to the tree topology of network foundation text
Poor sum or syntactic analysis are obtained) merge to carry out bottom-up semantic expressiveness recurrence, finally give the language of whole text
Justice is represented.At present, RecNN is used widely in many tasks of natural language processing, for example, emotional semantic classification [1],
Repeat detection [2], semantic analysis [3] and statistical machine translation [4] [5] etc..Self-encoding encoder (Auto-Encoder, abbreviation AE)
Initially proposed as a dimensionality reduction skill, now, AE more be used for obtain dimension it is higher and it is significant represent.
AE is made up of encoder and decoder two parts, and further decoding is exported after input is encoded, and is inputted by minimizing
Reconstructed error between output obtains more accurate and significant semantic expressiveness.At present, AE occurs by continuing to develop
Many mutation, such as Denoising Auto-Encoder (DAE), Constractive Auto-Encoder (CAE) etc..
In natural language processing, statistical method has in occupation of consequence in the tasks such as participle, part-of-speech tagging, syntactic analysis
Outstanding performance.But compared with the language phenomenon that statistical language can be described, the natural language in practical application will complexity
Much, particularly various constraintss.Graph model combines graph theory and statistical method, should the reasoning based on figure
Use in probability statistics framework, a kind of feasible thinking is proposed for various complicated restriction relations in description natural language.Mesh
Before, graph model is used widely [6] [7] [8] in natural language processing task, such as syntactic analysis.
In terms of the insertion of bilingual phrase represents research, traditional method [9] [10] [11] [12] mainly includes two steps:
1) the embedded expression of single language phrase is generated respectively using recurrence self-encoding encoder (Recursive Auto-Encoder, abbreviation RAE);
In the specific implementation, conventional method sets up the binary tree structure corresponding with phrase according to the minimum principle of reduction error sum, so
Semantic expressiveness recurrence merging is carried out based on the tree construction afterwards, the embedded of generation phrase represents vector;2) in bilingual recurrence certainly
Propose in encoder (Bilingual Auto-Encoder, abbreviation BRAE), according to bilingual phrase it is semantic identical the characteristics of come
Carry out the mutual supervised training that bilingual phrase insertion is represented.However, conventional method only consider in modeling process reconstructed error and
The semantic corresponding relation of bilingual phrase, and lack and consider more fully semantic constraint relation.Therefore, existing method is still deposited
In deficiency, how to learn to obtain preferably bilingual phrase insertion expression is still a good problem to study.
Bibliography:
[1]Richard Socher,Jeffrey Pennington,Eric H.Huang,Andrew Y.Ng,and
Christopher D.Manning.Semi-Supervised Recursive Autoencoders for Predicting
Sentiment Distributions.In Proc.Of EMNLP,2011.
[2]Richard Socher,Eric H.Huang,Jeffrey Pennin,Christopher D Manning,
and Andrew Y.Ng.Dynamic pooling and unfolding recursive autoencoders for
paraphrase detection.In Proc.Of NIPS,pages 801–809.2011.
[3]Richard Socher,Cliff Chiung-Yu Lin,Andrew Y.Ng,and Christopher
D.Manning.Parsing Natural Scenes and Natural Language with Recursive Neural
Networks.In Proc.of ICML,2011.
[4]Jiajun Zhang,Shujie Liu,Mu Li,Ming Zhou,and Chengqing
Zong.Bilingually-constrained phrase embeddings for machine translation.In
Proc.of ACL,pages111–121,June 2014.
[5]Jinsong Su,Deyi Xiong,Biao Zhang,Yang Liu,Junfeng Yao,and Min
Zhang.Bilingual correspondence recursive autoencoder for statistical machine
translation.In Proc.of EMNLP,pages1248–1258,September 2015.
[6]Andrei Alexandrescu and Katrin Kirchhoff.2009.Graph-based learning
for statistical machine translation.In Proc.of NAACL 2009,pages 119-127.
[7]Shujie Liu,Chi-Ho Li,Mu Li,and Ming Zhou.2012.Learning translation
consensus with structured label propagation.In Proc.of ACL 2012,pages 302-
310.
[8]Xiaoning Zhu,Zhongjun He,Hua Wu,Haifeng Wang,Conghui Zhu,and
Tiejun Zhao.2013.Improving pivot-based statistical machine translation using
random walk.In Proc.of EMNLP2013,pages 524-534.
[9]Jianfeng Gao,Xiaodong He,Wen-tau Yih,and Li Deng.2014.Learning
continuous phrase representation for translation modeling.In Proc.of ACL
2014,pages 699-709
[10]Kyunghyun Cho,Bart van Merrienber,Caglar Culcehre,Dzmitry
Bahdanau,Fethi Bougares,Holger Schwenk,and Yoshua Bengio.2014.Learning phrase
representations using rnn encoder-decoder for statistical machine
translation.In Proc.of EMNLP 2014,pages 1724-1734.
[11]Shixiang Lu,Zhenbiao Chen,and Bo Xu.2014.Learning new semi-
supervised deep auto-encoder features for statistical machine translation.In
Proc.of ACL 2014,pages 122-132.
[12]Xing Wang,Deyi Xiong,and Min Zhang.2015.Learning semantical
representations for nonterminals in hierarchical phrase-based translation.In
Proc.of EMNLP 2015,pages 1391-1400.
The content of the invention
It is an object of the invention to provide the bilingual recurrence self-encoding encoder based on figure.
The present invention comprises the following steps:
1) bilingual phrase is extracted from parallel corpora as training data, the translation probability between bilingual phrase is calculated;
2) method based on pivot, calculates and repeats probability;
3) the semantic relation figure of bilingual phrase is constructed;
In step 3) in, the specific method of the semantic relation figure of the bilingual phrase of construction can be:With source phrase and mesh
Mark end phrase is node, to any source phrase and destination end phrase, if belonging to the phrase pair in bilingual phrase language material, structure
Make a company side.All sets of node and side collection constitute the semantic relation figure of corresponding bilingual phrase;
4) the semantic relation figure based on bilingual phrase;
In step 4) in, the specific method of the semantic relation figure based on bilingual phrase can be:Define two kinds it is implicit
Semantic constraint condition, for two different nodes in same language, if they are connected to the same node point collection of another language
Close, then it is assumed that they are close to each other on semantic space, this is constraint one;For a kind of any node of language, its with it is another
The semantic correlation intensity of the different adjacent nodes of kind of language should be close with the phrase translation probability that is obtained based on maximal possibility estimation
Correlation, this is constraint two.
5) quantitative model object function, carries out model parameter training.
In step 5) in, the quantitative model object function, carrying out the specific method of model parameter training can be:Traditional
Bilingual recurrence self-encoding encoder includes single language reconstructed error and bilingual alignment uniformity fraction;On this basis, while by definition
Two implicit semantic constraint conditionings are on bilingual phrase, and the single language similar consistency fraction of introducing and bilingual translation are distributed one
Cause property fraction.
The present invention is expressed as target preferably to learn bilingual phrase insertion, lacks for conventional method and considers natural language
In more fully semantic constraint relation, it is proposed that a kind of bilingual recurrence self-encoding encoder based on figure.Inventive algorithm is clear and definite, think
Road is clear, and this method can improve the bilingual phrase insertion acquired and represent, preferably act on natural language processing task.
The present invention constructs the semantic relation figure of bilingual phrase first, and two implicit semantic constraints are defined by graph structure, uses
Represent, and then be preferably applied in natural language processing task in the more accurate bilingual phrase insertion of study, as machine is turned over
Translate.
The concretism of the present invention is as follows:
Graph model can often describe restriction relation more complicated in natural language.The present invention traditional bilingual recurrence from
On the basis of encoder, the semantic relation figure of bilingual phrase is constructed first, and more fully semantic knowledge is explored by graph structure,
And further define two implicit semantic constraints:For two different nodes in same language, if they are connected to separately
The same node point set of one language, then it is assumed that they are close to each other on semantic space, this is constraint one.For a kind of language
Any node, the semantic correlation intensity of its adjacent node different from another language should be with being obtained based on maximal possibility estimation
Phrase translation probability is closely related, and this is constraint two.Finally acted on the bilingual phrase of extraction, so as to learn to obtain more
Plus accurately bilingual phrase insertion is represented.
Brief description of the drawings
Fig. 1 is traditional BRAE model frameworks.
Fig. 2 is semantic relation graph structure example of the invention.In fig. 2, a represents the semantic relation of the bilingual phrase of construction
Figure, b represents the subgraph example of implicit semantic constraint one, and c represents the subgraph example of implicit semantic constraint two, vfNode represents source
Phrase, veNode represents destination end phrase, and solid black lines represent between connected source phrase and destination end phrase there is alignment to close
System, is used as a company side of semantic relation figure.
Embodiment
The specific embodiment of the invention is as follows:
The first step, extracts bilingual phrase as training data, the translation calculated between bilingual phrase is general from parallel corpora
Rate.
Second step, the method based on pivot calculates and repeats probability.
3rd step, constructs the semantic relation figure of bilingual phrase.Using source phrase and destination end phrase as node, to any source
Phrase and destination end phrase are held, if belonging to the phrase pair in bilingual phrase language material, a company side is constructed.All sets of node and
Side collection constitutes the semantic relation figure of corresponding bilingual phrase.
4th step, the semantic relation figure based on bilingual phrase defines two kinds of implicit semantic constraint conditions.For same language
The two different nodes called the turn, if they are connected to the same node point set of another language, then it is assumed that they are in semantic space
Upper close to each other, this is constraint one.For a kind of any node of language, the language of its adjacent node different from another language
Adopted correlation intensity should be closely related with the phrase translation probability obtained based on maximal possibility estimation, and this is constraint two.
5th step, quantitative model object function carries out model parameter training.Traditional bilingual recurrence self-encoding encoder is comprising single
Language reconstructed error and bilingual alignment uniformity fraction.On this basis, while by two of definition implicit semantic constraint conditions
Act on bilingual phrase, introduce single language similar consistency fraction and bilingual translation distribution uniformity fraction.
Committed step is described below realizes details:
1. the semantic relation figure of the bilingual phrase of construction
Graph model combines graph theory and statistical method, and the reasoning based on figure is applied in probability statistics framework, is
Various complicated restriction relations in description natural language propose a kind of feasible thinking.The present invention is proposed based on the bilingual of figure
Recurrence self-encoding encoder, it is necessary first to construct the semantic relation figure of bilingual phrase.Left figure a such as Fig. 2 is the semantic relation of construction
A part for figure.First, using all source phrase and destination end phrase as the node of figure, such as source phrase is " so remote
" so far away " are as the node of figure for (ruci yuan) " and destination end phrase.Then, for any source phrase and mesh
End phrase is marked, if belonging to the phrase pair in bilingual phrase language material, a company side, such as source phrase " so remote (ruci is constructed
Yuan) " and destination end phrase " so far away " are the alignment phrases in bilingual corpora, so node nf1And ne1Between have one
Bar connects side, and source phrase " so remote (ruci yuan) " and destination end phrase " up to now " and non-alignment phrase, so
Node nf1And ne4Between the company of having no side.
2. define two kinds of implicit constraint conditions
Traditional bilingual recurrence self-encoding encoder includes single language reconstructed error and bilingual alignment uniformity fraction, and this is based on aobvious
The semantic information of formula and realize, but this is inadequate for the Precise Representation for learning bilingual phrase, so the present invention proposes base
In the bilingual recurrence self-encoding encoder of figure, from graph theory, deeper implicit semantic constraint information is explored, two are thus defined
Individual implicit semantic constraint condition.
For two different nodes in same language, if they are connected to the same node point set of another language,
Think that they are close to each other on semantic space, this is constraint one.Such as right figure top b, source phrase node n in Fig. 2f2And nf3
It is connected to identical destination end phrase book { n simultaneouslye2,ne3,ne4, then think node nf2And nf3Closely should may be used on semantic space
Can it is close.The repetition probability for so calculating two source phrases with same target end phrase node collection is as follows:
Wherein, f, fsRepresent source phrase, e represents destination end phrase, p (f | e) and p (e | fs) represent based on maximum likelihood
Estimate the translation probability calculated.Then, definition is repeated to f, fsBetween weight it is as follows:
For any source phrase f, calculate all repetitions where it to associated weight, it is final only retain there is maximum
The repetition pair of weight.The calculating of weight between is repeated for destination end phrase to repeat to similar with source phrase.Afterwards, can be with
The error function of the single language similitude of definition:
Wherein,WithRepresent respectively source repeat pair and destination end repeat to Euclidean distance, θ represents
The parameter of model.
For a kind of any node of language, the semantic correlation intensity of its adjacent node different from another language should be with
The phrase translation probability obtained based on maximal possibility estimation is closely related, and this is constraint two.Such as Fig. 2 lower right figure c, source section
Point nf3It is connected to destination end set of node { ne2,ne3,ne4, embedded according to phrase is represented, n can be tried to achieve respectivelyf3To ne2, ne3, ne4
Semantic correlation intensity value, and estimate that the translation probability between them is distributed based on this, this probability distribution should be with
The probability distribution obtained based on maximal possibility estimation is tried one's best unanimously.Thus, it is possible to define the semantic correlation intensity point of adjacent node
Number is as follows:
Wherein, e, f represent destination end phrase and source phrase respectively,Respectively its corresponding embedded expression,
WithRespectively corresponding transposed matrix and deviation, tran (f) gather for f candidate's translation.Original language is portrayed with KL divergences
In the semantic correlation intensity fraction and based on maximal possibility estimation obtain translation distribution score between similitude, following institute
Show:
Etran(f, θ)=count (f) KL (p (* | f), psc(*|f))
Wherein, p (* | f) represents the translation probability obtained based on maximal possibility estimation, and count (f) represents f in training corpus
The number of times of middle appearance, θ is corresponding model parameter.The similarity scores of object language calculate similar, in this way, definable is based on
The error function of the uniformity of bilingual translation distribution is as follows:
Etran(nf,ne, θ) and=Etran(f,θ)+Etran(e,θ)
3. the object function and training method of model
The object function of the model of the present invention is mainly comprising four parts:Single language reconstructed error of traditional RAE models and
The bilingual alignment conformity error of BRAE models, and single language similitude based on the two implicit semantic constraints of the invention defined
Fraction and translation distribution uniformity fraction.It is specifically described as follows:
1) single language reconstructed error, the fraction has modeled the quality that phrase insertion is represented.First, by adjacent node generation father's section
In point, such as Fig. 1 dashed rectangles, phrase p=(x are inputted1,x2,x3), by x1And x2Generate corresponding father node as follows:
Wherein,Respectively x1、x2Term vector represent that f is activation primitive, here select tanh (), W(1)And b(1)Respectively corresponding weight matrix and straggling parameter.Then, by father nodeThe original child node of reconstruct is as follows:
Afterwards, by father nodeAnd child nodeNew father node is generated againSo constantly be recursively combined and
Reconstruct the expression until generating whole phrase.Thus, the quality that phrase insertion is represented is modeled according to reconstructed error:
Wherein, T (p) is the set of father node,For the expression of word initial in phrase,For in phrase
The expression of the word of reconstruct, cost function is the Euclidean distances for calculating two expressions.
2) bilingual alignment conformity error, the fraction has modeled loss function of the bilingual phrase in alignment uniformity.Such as
Shown on the right of Fig. 1 dashed rectangles, it is defined as follows:
Wherein, f and e are respectively source phrase and destination end phrase,WithIt is embedded in and represents for corresponding phrase,With
Respectively corresponding transposed matrix and deviation.
3) error function based on implicit semantic constraint, shown in its face as defined above.
The overall goals function J of modelGBRAEIt is defined as follows:
JGBRAE(nf,ne)=α Erec(nf,ne;θ)+β·Esem(nf,ne;θ)+g·(Esyn(nf,ne;θ)+Etran(nf,
ne;θ))
Wherein, a, β, g are super ginseng, for balancing every loss function, and alpha+beta+g=1.To avoid model over-fitting, draw
Enter parameter regularization term R (θ).Finally, the object function of model is as follows:
Wherein, G is all alignment phrase set, and N is the number of alignment phrase in training corpus, λ*For regularization term, R
(θ) is model parameter.
The object function not only allows for the explicit semantic constraint condition in traditional BRAE, and it is of the invention new fixed also to have incorporated
Two kinds of implicit semantic constraint conditions of justice, can make full use of various semantic constraint information, so as to learn more accurate bilingual short
Language insertion is represented, and is used in follow-up experiment, such as machine translation.
Claims (4)
1. the bilingual recurrence self-encoding encoder based on figure, it is characterised in that comprise the following steps:
1) bilingual phrase is extracted from parallel corpora as training data, the translation probability between bilingual phrase is calculated;
2) method based on pivot, calculates and repeats probability;
3) the semantic relation figure of bilingual phrase is constructed;
4) the semantic relation figure based on bilingual phrase;
5) quantitative model object function, carries out model parameter training.
2. the bilingual recurrence self-encoding encoder as claimed in claim 1 based on figure, it is characterised in that in step 3) in, the construction is double
The specific method of the semantic relation figure of language phrase is:Using source phrase and destination end phrase as node, to any source phrase and
Destination end phrase, if belonging to the phrase pair in bilingual phrase language material, constructs a company side.All sets of node and side collection are constituted
The semantic relation figure of corresponding bilingual phrase.
3. the bilingual recurrence self-encoding encoder as claimed in claim 1 based on figure, it is characterised in that in step 4) in, it is described based on double
The specific method of the semantic relation figure of language phrase is:Two kinds of implicit semantic constraint conditions are defined, for two in same language
Individual different nodes, if they are connected to the same node point set of another language, then it is assumed that they are close to each other on semantic space,
This is constraint one;For a kind of any node of language, the semantic correlation intensity of its adjacent node different from another language
It should be closely related with the phrase translation probability obtained based on maximal possibility estimation, this is constraint two.
4. the bilingual recurrence self-encoding encoder as claimed in claim 1 based on figure, it is characterised in that in step 5) in, the quantization mould
Type object function, carry out model parameter training specific method be:Traditional bilingual recurrence self-encoding encoder includes single language reconstruct and missed
Difference and bilingual alignment uniformity fraction;On this basis, while by two of definition implicit semantic constraint conditionings in double
On language phrase, single language similar consistency fraction and bilingual translation distribution uniformity fraction are introduced.
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201710257714.4A CN107092594B (en) | 2017-04-19 | 2017-04-19 | Bilingual recurrence self-encoding encoder based on figure |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201710257714.4A CN107092594B (en) | 2017-04-19 | 2017-04-19 | Bilingual recurrence self-encoding encoder based on figure |
Publications (2)
Publication Number | Publication Date |
---|---|
CN107092594A true CN107092594A (en) | 2017-08-25 |
CN107092594B CN107092594B (en) | 2019-07-09 |
Family
ID=59637832
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN201710257714.4A Expired - Fee Related CN107092594B (en) | 2017-04-19 | 2017-04-19 | Bilingual recurrence self-encoding encoder based on figure |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN107092594B (en) |
Cited By (6)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN108363704A (en) * | 2018-03-02 | 2018-08-03 | 北京理工大学 | A kind of neural network machine translation corpus expansion method based on statistics phrase table |
CN110334358A (en) * | 2019-04-28 | 2019-10-15 | 厦门大学 | A kind of phrase table dendrography learning method of context-aware |
CN110472252A (en) * | 2019-08-15 | 2019-11-19 | 昆明理工大学 | The method of the more neural machine translation of the Chinese based on transfer learning |
CN110489752A (en) * | 2019-08-14 | 2019-11-22 | 梁冰 | A kind of semantic recurrence expression system of natural language |
CN111949792A (en) * | 2020-08-13 | 2020-11-17 | 电子科技大学 | Medicine relation extraction method based on deep learning |
WO2023071123A1 (en) * | 2021-10-29 | 2023-05-04 | 广东坚美铝型材厂(集团)有限公司 | Self-learning method for semantic feature with maximum gap, and computer device and storage medium |
Citations (4)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN103559181A (en) * | 2013-11-14 | 2014-02-05 | 苏州大学 | Establishment method and system for bilingual semantic relation classification model |
CN103577399A (en) * | 2013-11-05 | 2014-02-12 | 北京百度网讯科技有限公司 | Method and device for extension of data in bilingual corpuses |
CN104268132A (en) * | 2014-09-11 | 2015-01-07 | 北京交通大学 | Machine translation method and system |
CN105005554A (en) * | 2015-06-30 | 2015-10-28 | 北京信息科技大学 | Method for calculating word semantic relevancy |
-
2017
- 2017-04-19 CN CN201710257714.4A patent/CN107092594B/en not_active Expired - Fee Related
Patent Citations (4)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN103577399A (en) * | 2013-11-05 | 2014-02-12 | 北京百度网讯科技有限公司 | Method and device for extension of data in bilingual corpuses |
CN103559181A (en) * | 2013-11-14 | 2014-02-05 | 苏州大学 | Establishment method and system for bilingual semantic relation classification model |
CN104268132A (en) * | 2014-09-11 | 2015-01-07 | 北京交通大学 | Machine translation method and system |
CN105005554A (en) * | 2015-06-30 | 2015-10-28 | 北京信息科技大学 | Method for calculating word semantic relevancy |
Non-Patent Citations (1)
Title |
---|
苏劲松等: "引入基于主题复述知识的统计机器翻译模型", 《浙江大学学报(工学版)》 * |
Cited By (9)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN108363704A (en) * | 2018-03-02 | 2018-08-03 | 北京理工大学 | A kind of neural network machine translation corpus expansion method based on statistics phrase table |
CN110334358A (en) * | 2019-04-28 | 2019-10-15 | 厦门大学 | A kind of phrase table dendrography learning method of context-aware |
CN110489752A (en) * | 2019-08-14 | 2019-11-22 | 梁冰 | A kind of semantic recurrence expression system of natural language |
CN110489752B (en) * | 2019-08-14 | 2021-06-22 | 梁冰 | Semantic recursion representation system of natural language |
CN110472252A (en) * | 2019-08-15 | 2019-11-19 | 昆明理工大学 | The method of the more neural machine translation of the Chinese based on transfer learning |
CN110472252B (en) * | 2019-08-15 | 2022-12-13 | 昆明理工大学 | Method for translating Hanyue neural machine based on transfer learning |
CN111949792A (en) * | 2020-08-13 | 2020-11-17 | 电子科技大学 | Medicine relation extraction method based on deep learning |
CN111949792B (en) * | 2020-08-13 | 2022-05-31 | 电子科技大学 | Medicine relation extraction method based on deep learning |
WO2023071123A1 (en) * | 2021-10-29 | 2023-05-04 | 广东坚美铝型材厂(集团)有限公司 | Self-learning method for semantic feature with maximum gap, and computer device and storage medium |
Also Published As
Publication number | Publication date |
---|---|
CN107092594B (en) | 2019-07-09 |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
CN107092594B (en) | Bilingual recurrence self-encoding encoder based on figure | |
CN106202010B (en) | Method and apparatus based on deep neural network building Law Text syntax tree | |
Chen et al. | Zero-shot text classification via knowledge graph embedding for social media data | |
Tang et al. | Deep sequential fusion LSTM network for image description | |
Zeng et al. | Conceptual and syntactical cross-modal alignment with cross-level consistency for image-text matching | |
CN113779220A (en) | Mongolian multi-hop question-answering method based on three-channel cognitive map and graph attention network | |
CN110427616A (en) | A kind of text emotion analysis method based on deep learning | |
Zeng et al. | Aspect-level sentiment analysis based on semantic heterogeneous graph convolutional network | |
Liu et al. | Enhancing aspect-based sentiment analysis using a dual-gated graph convolutional network via contextual affective knowledge | |
Lu et al. | Construction of multimedia assisted legal classroom teaching model based on data mining algorithm | |
Xu et al. | A novel entity joint annotation relation extraction model | |
Wang et al. | Fusion layer attention for image-text matching | |
Shang et al. | Aspect-level sentiment analysis based on aspect-sentence graph convolution network | |
Krause et al. | Sar-graphs: A linked linguistic knowledge resource connecting facts with language | |
Xia et al. | Generating Questions Based on Semi-Automated and End-to-End Neural Network. | |
CN113392220A (en) | Knowledge graph generation method and device, computer equipment and storage medium | |
Zhao et al. | Fusion with GCN and SE-ResNeXt network for aspect based multimodal sentiment analysis | |
Hua et al. | A semantic hierarchical graph neural network for text classification | |
Wu et al. | Multi-hop community question answering based on multi-aspect heterogeneous graph | |
Li | Automatic semantic analysis framework of Dickinson’s portfolio based on character recognition and artificial intelligence | |
Xu et al. | BERT-PAGG: a Chinese relationship extraction model fusing PAGG and entity location information | |
Hu et al. | Exploring discourse structure in document-level machine translation | |
Jiang et al. | Constructing semantic hierarchies via fusion learning architecture | |
Wang | Research on spatial conceptual modeling of natural language processing based on deep learning algorithms | |
Yu et al. | SSAG-Net: Syntactic and Semantic Attention-Guided Machine Reading Comprehension. |
Legal Events
Date | Code | Title | Description |
---|---|---|---|
PB01 | Publication | ||
PB01 | Publication | ||
SE01 | Entry into force of request for substantive examination | ||
SE01 | Entry into force of request for substantive examination | ||
GR01 | Patent grant | ||
GR01 | Patent grant | ||
CF01 | Termination of patent right due to non-payment of annual fee | ||
CF01 | Termination of patent right due to non-payment of annual fee |
Granted publication date: 20190709 Termination date: 20200419 |